Whisper large-v3 — Mongolian Speech-to-Text Benchmark
Real results from the Whisper large-v3 Batch API (enhanced operating point, language mn) evaluated across 4 Mongolian datasets and 265 audio samples. Every number below is measured — not marketing. WER, speed, and pricing are shown as-is, brutally honest.
265/265 samples transcribed · 100% success rate
Lower is better · across 265 samples
Pricing
Whisper large-v3 list pricing for batch transcription. No discounts, no negotiated rates applied — the raw per-minute rate.
Pricing source: Whisper large-v3 public pricing. Duudlaga Flow is shown for context only — this page isolates Whisper large-v3 so the number is not padded by our own product.
WER & CER by dataset
Word and character error rates per dataset. Lower is better — and these are the real Whisper large-v3 numbers, which are weak on Common Voice 24.
Dataset summary
Aggregate accuracy, speed, and timing for each dataset.
| Dataset | Source | Samples | Success | WER | CER | Accuracy | Speed | Audio (s) | Proc (s) |
|---|---|---|---|---|---|---|---|---|---|
| Common Voice 24 (MN) | Hugging Face → | 59 | 59/59 | 91.2% | 38.8% | 8.8% | 0.72× | 330.6s | 456.5s |
| Shunya Labs Mongolian Speech | Hugging Face → | 60 | 60/60 | 86.2% | 33% | 13.8% | 1.87× | 645.1s | 344.8s |
| Common Voice 20 (MN) | Hugging Face → | 54 | 54/54 | 94% | 41% | 6.0% | 1.13× | 269.8s | 238.3s |
| Modern Voice | — | 92 | 92/92 | 87.7% | 37.2% | 12.3% | 1.53× | 603.8s | 394.1s |
WER vs speed — per sample
Each dot is one audio sample. The sweet spot is the bottom-left (low error, fast). Whisper large-v3 sits high on error for many Common Voice 24 samples.
Per-sample results
Ground truth shown verbatim in the Expected column. The result column highlights only the words Whisper large-v3 got wrong, in red — no strikethrough/swap gymnastics, just the mistakes.
| # | Audio | Sample | Dataset | Expected (ground truth) | Whisper large-v3 result | WER | CER | I/D/S |
|---|---|---|---|---|---|---|---|---|
| 1 | btsee_0001 | Common Voice 24 (MN) | Гэхдээ амьсгал хураахаасаа өмнө танд мэдэж байгаагаа хэлье. | Кэлтээ амсгол түрээ хасаа омон танд мэдэж байгаа хэлээ. | 87.5% | 32.8% | 1/0/6 | |
| 2 | btsee_0002 | Common Voice 24 (MN) | Надад заяасан аз жаргал гэдэг ердөө гуравхан сарын хугацаатай байсан гэж үү? | Надзай асан аджирал гэдэгээрдээ үрхүн сарай хүгцадтай байсан гэж бүх. | 83.3% | 33.3% | 0/2/8 | |
| 3 | btsee_0003 | Common Voice 24 (MN) | Одоо бид өөрсдөө өвчин эмгэгээсээ салахыг хичээцгээе. | Атаах тээ төрч төрч энрээгээсээ салхих чигцгээ. | 100.0% | 53.8% | 0/0/7 | |
| 4 | btsee_0004 | Common Voice 24 (MN) | Би бол голдуу хээрээр гэр, хэцээр дэр хийж явдаг хүн. | Би босголту хэрээр гэр хэцэр дэрхийж ябдог хүн. | 70.0% | 17.6% | 0/2/5 | |
| 5 | btsee_0005 | Common Voice 24 (MN) | Хан хурмаст уурлаж, Болдоггүй Бор өвгөнийг хор луугаараа ниргүүлэхээр явуулжээ. | Ханхурмас дуу орлос болдгой бараа үгэнээ 2 лоогуу арны эргүйлэхээр ягалж байна. | 120.0% | 51.9% | 2/0/10 | |
| 6 | btsee_0006 | Common Voice 24 (MN) | Алив наашаа ороод ир гээд гэртээ оров. | Аж байнааша ородор, гэдгэжтэй оров. | 85.7% | 35.1% | 0/2/4 | |
| 7 | btsee_0007 | Common Voice 24 (MN) | Өө өндөр дээдэс таны тухайд би баталж чадахгүй. | Уу, үндэр дээд таныи тухууд би бааталг чалтахгүй. | 87.5% | 28.3% | 0/0/7 | |
| 8 | btsee_0008 | Common Voice 24 (MN) | Харин гурав дахь удаагаас эхлэн хүмүүсийг сонирхож эхлэв. | Харин 3-тэг хуудагаса эхлэн хүмүүстэг сайнархаж гэлэв. | 75.0% | 35.7% | 0/0/6 | |
| 9 | btsee_0009 | Common Voice 24 (MN) | Та нар очингуутаа шөл л өгч үз. | Танр очингууд шүүллэл үхж байз. | 100.0% | 50.0% | 0/2/5 | |
| 10 | btsee_0010 | Common Voice 24 (MN) | Ерөөсөө литр үйлдвэрээсээ салаагүй явсан юм чинь. | Ерөөсөөл үүлтүрөссээ салаагуу ябсан имч нь. | 100.0% | 33.3% | 0/1/6 |
Methodology
How these numbers were produced.
Provider: Whisper large-v3 (OpenAI Whisper large-v3 open weights run locally with an explicit Mongolian language token — the reference open-source ASR baseline.).
Endpoint local (transformers). Language mn. Operating point openai/whisper-large-v3 (language=mn, task=transcribe) on mps. Diarization none.
Datasets: Common Voice 24 (MN), Shunya Labs Mongolian Speech, Common Voice 20 (MN), Modern Voice — 265 samples, 1849.3s of audio total.
- Common Voice 24 (MN)https://huggingface.co/datasets/btsee/common-voices-24-mn
- Shunya Labs Mongolian Speechhttps://huggingface.co/datasets/shunyalabs/mongolian-speech-dataset
- Common Voice 20 (MN)https://huggingface.co/datasets/warmestman/common-voice-20-mn-normalized
- Modern Voice—
Metrics: WER and CER are computed with a standard word/character Levenshtein alignment, normalized for case and punctuation. Accuracy = 100 − WER. Speed factor = audio duration ÷ processing time (× realtime). All requests are real Whisper large-v3 Batch API calls, not cached or simulated.
Diff highlighting: The result column aligns to the ground truth and colors every substitution and insertion red. Deletions (words missing from the result) are not shown in the result column — the Expected column already holds the full ground truth as-is.